bioRxiv preprint doi: https://doi.org/10.1101/499392; this version posted May 14, 2019. The copyright holder for this preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. A Modeling Framework for Exploring Sampling and Observation Process Biases in Genome and Phenome-wide Association Studies using Electronic Health Records Lauren J. Beesley∗1, Lars G. Fritsche1, and Bhramar Mukherjee1 1University of Michigan, Department of Biostatistics *Corresponding Author: [email protected] May 14, 2019 Abstract Large-scale association analyses based on observational health care databases such as electronic health records have been a topic of increasing interest in the scientific community. However, challenges of non-probability sampling and phenotype misclassification associated with the use of these data sources are often ignored in standard analyses. In general, the ex- tent of the bias that may be introduced by ignoring these factors is not well-characterized. In this paper, we develop a statistical framework for characterizing the bias expected in associ- ation studies based on electronic health records when disease status misclassification and the sampling mechanism are ignored. Through a sensitivity analysis approach, this framework can be used to obtain plausible values for parameters of interest given results obtained from standard na¨ıve analysis methods. We develop an online tool for performing this sensitivity analysis. Simulations demonstrate promising properties of the proposed approximations. We apply our approach to study bias in genetic association studies using electronic health record data from the Michigan Genomics Initiative, a longitudinal biorepository effort within Michigan Medicine. Keywords: non-probability sampling, electronic health records, outcome misclassification, Phe- WAS, GWAS 1 Introduction Genome-wide genotype data linked with electronic health records (EHR) are becoming increas- ingly available through biorepository efforts at academic medical centers, health care organi- zations, and population-based biobanks [1]. A common use of these linked data is to explore the association between a phenotype, D, with a risk factor of interest, G, after adjusting for potential confounders, Z. Analysis using a regression model for DjG; Z may be repeated for millions of risk factors or genetic variants with a given D of interest (as in a genome-wide association study (GWAS)) or for thousands of phenotypes derived from the content of the EHR with a given variant G of interest (as in phenome-wide association studies (PheWAS)). Association analyses embedded within large observational databases have gained popularity in recent years, and the use of and interest in such analyses continues to increase [1,2, 3]. However, unlike curated and well-designed population-based studies, data obtained from large observational databases are often not originally intended for research purposes, and additional thought is needed to understand potential sources of bias. In this paper, we focus on the par- ticular association setting with D being a single EHR-derived phenotype and G being a single genetic marker or a polygenic risk score, but the methods and conceptual framework developed in this paper can be applied quite broadly to general estimation problems using observational 1 bioRxiv preprint doi: https://doi.org/10.1101/499392; this version posted May 14, 2019. The copyright holder for this preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. databases. One potential source of bias in EHR-based studies is misclassification of derived disease status variables. Large-scale agnostic studies using EHR often define patient disease status (phenotypes) based on international classification of disease (ICD) codes or aggregates thereof called \PheWAS codes" or \phecodes" to define phenotypes in an automated and reproducible way [4]. In practice, EHR-derived phenotype definitions are used to represent the underlying `true' disease status. However, these ICD-based phenotype classifications can be erroneous in capturing the `true' disease status for a variety of reasons. For example, psychiatric diseases can be difficult to diagnose, and diagnosis can often be subjective [5]. ICD code-based diagnoses may be an incomplete representation of a patient's health state, which may also be recorded in doctor's notes and elsewhere in the EHR. In response to this problem, there exists an extensive literature on using other structured and unstructured content of the EHR to define phenotypes more accurately [6, 7, 8, 9, 10]. Additionally, human validation can be used to evaluate phe- notyping algorithms [11]. These existing phenotyping approaches can be effective in reducing misclassification given the information available in the EHR, but even the most sophisticated phenotyping algorithms cannot capture diagnoses that were never recorded in any form in the EHR. Secondary conditions may not always be entered into the EHR, and symptoms occurring between visits may not always be reported. The EHR cannot adequately capture diseases that a patient had prior to entry into the EHR (outside the observation window). The chance of correctly capturing a disease is inherently dependent on the length of stay in the EHR or the observation/encounter process for a given patient. We often have a systematic source of mis- classification (that we will call \observation window bias") due to a lack of comprehensiveness of the EHR in capturing diagnoses or medical care obtained from outside sources (e.g. at another health care center). Together, these various factors can lead to a potentially large degree of misclassification, particularly due to underreporting of disease. Several authors have proposed statistical methods for addressing misclassification of bi- nary phenotypes in EHR-based studies. The extent of misclassification can be described using quantities such as sensitivity and specificity, but these quantities can vary from population to population and from phenotype to phenotype [12]. Huang et al. [13] proposes a likelihood- based approach that integrates over unknown sensitivity and specificity values but requires some limited prior information about the sensitivity and specificity. Wang et al. [14] proposes an approach for incorporating both human-validated labels and error-prone phenotypes into the estimation, but this approach will not account for observation window bias. Duffy et al. [15] and Sinnott et al. [16] expand on results in the measurement error literature to relate parameters in the model for the true outcome with the model for the misclassified outcome, but Duffy et al. [15] focuses on outcome misclassification with binary risk factors, and Sinnott et al. [16] focuses on the setting in which the probability of having observed disease is explicitly modeled using a variety of information in the EHR. Additionally, all of these methods do not directly address the sampling mechanism. In addition to potential bias due to misclassification of disease phenotypes, the mechanism by which subjects are selected into the dataset can sometimes result in biased inference when not handled appropriately. Complex sampling designs in an epidemiologic study can be addressed using survey design techniques if the sampling strategy is known. However, the probability mechanism for inclusion of a subject into a biorepository is not a priori fixed or defined. In- teractions with the health care system are generated by the patient, and it can be difficult to understand the mechanism driving sampling as well as self-selection for donating biosamples, which may be related to a broad spectrum of patient factors including overall health. Several authors recommend adjusting for factors such as number of health care visits or referral status to better account for the sampling mechanism [17, 18]. Additionally, there is a belief in the lit- erature that gene-related association study results may be less susceptible to bias resulting from patient selection [19]. This belief stems from the assumption that an individual genetic locus is 2 bioRxiv preprint doi: https://doi.org/10.1101/499392; this version posted May 14, 2019. The copyright holder for this preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. not usually appreciably related to selection. However, bias due to genotype relationships with selection can still arise in certain settings [20]. Additionally, a popular topic in genetics-related research right now is the use of polygenic risk scores, which combines information from many genetic loci into a score to quantify a patient's genetic risk for developing a particular disease [21, 22]. While it may be reasonable to assume that a specific genetic locus may have little association with selection, this assumption becomes more tenuous for an aggregate polygenic risk score with stronger association with the underlying disease and other factors related to selection. As we will demonstrate, patient sampling can create substantial bias in estimating genetic associations using EHR data in the presence of disease status misclassification. Existing statis- tical methods for dealing with phenotype misclassification do not directly take into account the mechanism by which patients are sampled and vice versa. Additionally, standard association studies often do not account for either source